Plugging in your own model server

Mnemosyne can call a model server you own instead of a provider. Any server compatible with the OpenAI API works: vLLM, Ollama, TGI, or an internal gateway. The card is called "Your own server", in the Intelligence settings.

The two fields that matter

The base URL is your server's address. It comes from you and from nowhere else. It needs the scheme: http://host:port or https://host:port. On a private network http:// is perfectly fine and is not refused. An address without a scheme is rejected rather than guessed, because guessing would silently call the wrong machine.

The model name is free text. It is the exact name your server answers to, and it is not translated, completed, or replaced by a default. There is no list to pick from: nobody but you knows what your server hosts.

The API key is optional. Many internal servers ask for none, and when you give none, no authorization header is sent.

Checking before relying on it

The card shows the address that is actually called, computed by the same function the answers use. So you can see whether the /v1 you added, or left out, produces the call you think it does.

The "Check" button queries the model list at that same address and tells you what it found. A server that answers but exposes no catalogue is not declared broken: that is a gateway which does not publish its list, not a failure.

What this route does not do

It does not handle images. A visual attachment is refused upfront rather than sent to a server nobody knows can read it.

It does not display a cost. Mnemosyne does not know the bill for your own machine, so it does not invent a figure and does not say "free" either.